GROUNDWATER MODELING IN ALPINE KARST SYSTEMS: A MODEL ENSEMBLE GENERATOR TO EXPLORE STRUCTURAL UNCERTAINTY

Groundwater flow patterns in alpine systems are poorly understood, yet alpine groundwater systems may be primary sources of recharge and surface flows to regional systems. Karst aquifers are characterized by high-conductivity conduits embedded in a low-conductivity fractured matrix, resulting in extreme heterogeneity and variable groundwater flow behavior. The conduit network controls groundwater flow, but is often unmapped, making it difficult to apply numerical models to predict system behavior. We present a multi-model ensemble method to represent structural and conceptual uncertainty inherent in simulations of these systems with limited spatial information. We test the new method by applying it to a well-mapped, geologically complex long-term study site: the Gottesacker alpine karst system. Our ensemble generation process consists of three steps, linking existing tools: creating 3D geologic models using GemPy (a Python package), generating multiple conduit networks constrained by the geology using the Stochastic Karst Simulator (a MATLAB script), and finally running multiple flow simulations with different parameter sets through each network using the Storm Water Management Model (C-based software). Non-behavioral models are rejected based on the fit of the simulated spring discharge to the observed discharge. This approach captures a diversity of plausible system configurations and behaviors using minimal initial data and devoting minimal computational resources to parameter estimation. The ensemble can then be used to explore the relative importance of conduit network structure and hydraulic flow parameters. For the ensemble generated in this study, the network structure was more determinant of flow behavior than the hydraulic parameters, but multiple different structures yielded similar fits to the observed flow behavior. This suggests that while modeling multiple structures is important to adequately represent structural uncertainty, additional types of data are needed to discriminate among structures.